The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents an attack surface for indirect prompt injection as its primary purpose is to teach the construction of agentic loops that process untrusted user input and external tool results (e.g., API responses). While the examples do not include explicit sanitization or boundary markers, this is typical for simplified educational code snippets, and the risk is inherent to the function-calling architecture described.
Ingestion points: User input interpolation in the messages array and tool result ingestion via the tool_result content block in SKILL.md.
Boundary markers: The implementation uses the structured message format (roles 'user' and 'assistant') which provides native platform boundaries, though additional delimiters are not used.
Capability inventory: The skill header requests access to Read, Write, Edit, Bash(npm:*), and Grep, which are powerful capabilities; however, these are not exercised maliciously within the provided instruction code.
Sanitization: No explicit sanitization of tool outputs is demonstrated in the examples.
[EXTERNAL_DOWNLOADS]: The skill references the official Anthropic SDKs (@claude-ai/sdk for Node.js and anthropic for Python) and points to official documentation at docs.anthropic.com. These are legitimate and safe references for the stated purpose of the skill.

clade-embeddings-search